The disclosure obligation for AI-written text falls away as soon as a human has reviewed it on its substance and someone carries editorial responsibility. No label required, provided there is a name behind it — and that exposes exactly where the legislator places responsibility.
By Anthony Raaijmakers
There is an exemption in the AI Act that I had to read twice. The disclosure obligation for AI-written text falls away as soon as a human has reviewed that text on its substance and someone carries editorial responsibility for it. No label required, provided there is a name behind it. That sounds like a loophole, but it is exactly how the rule is meant to work, and it exposes where the legislator believes responsibility belongs.
Since 2 August 2026, the transparency rules in Article 50 of the AI Act have applied. From that moment on, it must be clear when someone is talking to an AI system or looking at AI-generated content. For business owners this is no longer an abstract question, because the rules affect everyone who runs a chatbot on their site or sends generated text and images out into the world.
The confusion I run into most often sits in two words that both appear to mean “marking”. The European Commission uses them deliberately apart.
Marking is the invisible mark in the output itself. That is an obligation of the provider, the party supplying the AI system. It concerns a machine-readable signal in text, image, video or audio, plus a means of detecting it. You do not see it and you do nothing for it.
Labelling is the visible label for the reader or viewer. That is your obligation when you deploy a system under your own name. And that obligation is narrow: it applies to deepfakes and to AI-written text on matters of public interest that nobody has reviewed editorially.
Practical consequence: your product copy and your customer service chat fall outside the labelling obligation, but whatever your supplier puts into that output happens regardless.
That human review is the hinge. The law is not aimed at AI, but at content on societal matters for which nobody is answerable. As soon as a human carries editorial responsibility, there is a name you can turn to again, and precisely the problem the provision addresses disappears. It is the same logic as an editor-in-chief who signs off on what goes into the newspaper.
Do note what the text actually says: there must have been a process of human review, and someone who carries the responsibility. Your name under a publication settles the second. If you push a generated text out unread, the first has not been met.
Anthropic has signed the code of practice under Article 50(2) and describes in its own help centre how the marking works: an invisible watermark in generated text, and signed provenance data following the C2PA standard in files. Models released after 2 August 2026 do this from day one; for older models the transition period is still running.
The most interesting part sits in the limitations they set out themselves. A watermark that is found does not prove who the author is, because AI may only have edited or translated. And a missing watermark does not prove that no AI was involved, because a screenshot or a format conversion strips provenance data straight out.
A page like that is the benchmark. Ask every provider in your chain where their version of it is.
The fines for breaching Article 50 are set out in Article 99(4) of the Regulation: up to 15 million euro or 3 percent of worldwide annual turnover, whichever is higher. For SMEs and start-ups, paragraph 6 reverses that: there, the lower of the two applies. At five million in turnover you are therefore talking about a ceiling of one hundred and fifty thousand, not 15 million. That does not make it unimportant, but it does make the headline a good deal less applicable to most Dutch companies.
The law and the supplier arrive at the same point, each from the other side. The technology provides the signal, the signature carries the responsibility. We build software with AI-amplified engineers, and this is exactly why we do not treat human review as a leftover step but as a fixed one. Not because the law asks for it, but because it is the only place where someone can actually sign off on the outcome.
Make a single list of every point where AI output leaves your organisation: the customer service chat, generated product copy, social visuals, summaries in your client portal. Add two columns: who is the provider, and who carries editorial responsibility. That second column removes more rows from your list than you expect, and for the rows that remain, you at least know exactly where you stand.
If you get stuck on which of your systems fall under which role, we are happy to take a look with you.